@atelier/workspace

general MCP Server

Atelier — autonomous design agent for the Google for Startups AI Agents Challenge 2026. Stitch generates. Atelier asks, converges, learns.

Verified
generalgeneralgcp
4 views2 stars0 forksv0.2.0-alphaApache-2.0

Why This Matters

Discovered via github-topic:mcp and last synced 3mo ago.

Verified
Source
github-topic:mcp
Stars
2
Last synced
3mo ago
Install
Check source

Install

Install instructions not detected yet

Check the source repository for the latest setup steps.

View source instructions
27
Tools
0
Resources
0
Prompts
Standard I/O
Transport

Available Tools (27)

Governance

None

Observability

None

N6

CSC-D — Constitutional Self-Critique for Design

Resource

URL

Architecture

Single LLM pass

N10

Convergence Spec RFC

Regenerate-not-edit

Interactive WYSIWYG refinement; high visual fidelity

N2

DEMAS-D — Per-axis Provenance Matrix Design Judge

N11

Public Eval Harness

Opaque

Session-scoped, not user-legible

N3

PerJudge — Per-Project DPO Judge with Hebbian Mutator

Prompt-literal

Proactive; reads codebase; exports HTML/Canva/PPTX; handoff bundle to Claude Code

Contribution

Status

N9

Open Eval Adapters Library

None

None

N5

EvoDesign — AlphaEvolve-Inspired K-Candidate Search

N14

WRAI — Web-Research-Augmented Intake

Cloud-native

Vendor-neutral

N8

Public Judge Calibration Dashboard

N4

PADI — Project-Agnostic Descriptor Inference

N13

PIP — Pre-Generation Intake Protocol

N7

Governed A2UI — built on A2UI v0.9 (server-to-client subset), flag-gated

Claim

Evidence

N15

MJG — Multi-Judge Governance

N12

RLRD — Recursive Long-Running Discipline

Live

--- ## Architecture Full system diagram: [architecture-diagram.md](docs/architecture/architecture-diagram.md) ```mermaid graph TB subgraph "User Layer" U["User / Designer"] DASH["Firebase Hosting<br/>Bench + Replay + Auth"] end subgraph "Auth Layer" FB["Firebase Auth<br/>Google SSO"] IAP["Cloud IAP"] end subgraph "API — Cloud Run" API["FastAPI<br/>POST /v1/generate"] end subgraph "Pipeline — 8-Node DAG" N1["N1: Brief Parser"] N14["N14: WRAI"] N2["N2: Source Resolver"] N3A["N3a: DDLC Specialist Pipeline<br/>6 roles via Stitch MCP"] N3C["N3c: Deterministic Gates<br/>6 axes"] N3D["N3d: Consensus Judge<br/>D-O-R-A-V"] N4["N4: Best Pick<br/>κ=0.70"] end subgraph "DPO Flywheel" TR["Trajectory Recorder"] MINER["DPO Pair Miner"] TUNER["GeneratorTuner"] VERTEX["Vertex AI Tuning"] end U --> DASH --> FB --> IAP --> API --> N1 N1 --> N14 --> N2 --> N3A --> N3C --> N3D --> N4 N4 --> TR --> BQ[("BigQuery")] MINER --> BQ TUNER --> MINER TUNER --> VERTEX --> N3A ``` --- ## Why Atelier? The #1 complaint across design-generation tools is not raw capability — it is legibility, control, and trust. Users cannot tell why an output was accepted, cannot see what tokens burned, cannot edit without regenerating from scratch, and get designs that drift off their brand. Invisible token burn is the top usage complaint across Stitch, v0, and Lovable. ### Comparison

N1

DGF-D2C — Deterministic-Gate-First Design-to-Convergence